COVER STORY Korea's 2-track AI strategy casts uncertainty over sovereign model project
Gov't urged to clarify distinction between new frontier AI, ongoing program
The government has been working to develop a national artificial intelligence (AI) foundation model, while planning to launch a separate effort to develop frontier models. This two-track strategy is raising questions: Why is the government pushing the new project and how will it draw the line between the two programs?
The Ministry of Science and ICT confirmed that the ongoing AI foundation model project will proceed with a third-stage evaluation, with two finalists expected to be selected in February. Separately, the government plans to allocate 4.7 trillion won ($3.5 billion) for a frontier AI program set to begin in March 2027.
The split reflects the government's growing concern that the rapid advances in global AI have made the foundation model project's original goals harder to achieve, while abandoning the project could undermine the ecosystem and capabilities built up over the past year.
Deputy Prime Minister and Science Minister Bae Kyung-hoon has repeatedly rejected speculations that the foundation model project is being scrapped.
“The national AI foundation model project has been pursued with the goal of expanding AI across the industrial and public sectors, increasing its use and developing a world-class model,” Bae said at the National Assembly’s annual audit in Seoul on Tuesday.
He emphasized the third-stage evaluation is under way and that the government aims to bring the models closer to the level of open-weight AI models used globally.
The project, lauched last year, had originally set a benchmark of reaching about 95 percent of the performance of leading global models. However, the government has since acknowledged that the AI landscape has changed dramatically, particularly after the emergence of Claude Mythos underscored the rapid pace of advances in AI capabilities.
"We believe an additional frontier AI project is needed to address challenges facing humanity that could not previously be solved and respond to cybersecurity threats, so the related project is being added,” Bae said, explaining that each of the two tracks will serve different needs.
Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon, front row second from right, watches a dancing robot at a physical artificial intelligence forum at the Korea Development Bank headquarters in Yeongdeungpo District, Seoul, Sept. 22. Yonhap
However, the separate allocation of resources to frontier AI is raising concerns about the role of the existing project and whether its two winners will continue to receive the support originally promised.
Experts questioned whether the two programs can maintain distinct roles, particularly if the new frontier AI program succeeds in developing a more capable model.
“Even if we pursue both tracks, once a frontier-level model emerges, domestic companies or public institutions are likely to use the frontier model rather than the national AI foundation model,” said Lee Seong-yeob, a professor at Korea University's Graduate School of Management of Technology.
“If that happens, we ultimately move toward the frontier model, so I wonder whether there is really a need to continue the national AI foundation model project just to comply with the policy.”
Lee said the competition had helped advance AI technology, but the original goal of the national AI foundation model project was becoming harder to achieve as leading global AI models continue to advance, suggesting that concentrating resources on frontier AI could be more effective.
Choi Kyoung-jin, director of the Center for AI, Data and Policy at Gachon University, emphasized the need for the government to first clearly define the goals of the frontier model and how long it expects the development process to take.
“When the national AI foundation model project first started, its purpose was actually to develop a frontier model at the time … But today's frontier and next year's frontier are different,” Choi said. “The government may launch a new project, but I think it needs to clearly present what it aims to achieve.”
LG AI Research co-President Lee Hong-lak presents EXAONE 4.0 during the LG AI Talk Concert event at LG Science Park in western Seoul, July 22, 2025. Courtesy of LG AI Research
Choi suggested that the government consider building on existing models rather than starting from scratch.
“It may actually be faster to upgrade existing models rather than develop something new from scratch,” he said, cautioning against duplicating spending on data and other resources if the two programs use the same datasets.
He also called for better use of the capabilities and models built up through the past year of competition under the national AI project.
“A model does not stop being useful just because it ranks lower; lightweight models have their own uses by service area, while frontier-level models have their own uses,” he said, arguing that the country needs capabilities across frontier, lightweight and industry-specific models.
“What Korea needs is the ability to supply models for diverse fields, rather than necessarily having only a frontier model.”
For the three teams still competing in the existing project — LG AI Research, SK Telecom and Upstage — the immediate question is how the project will evolve after the final-stage evaluation.
The government has said the two finalists will continue to receive support, but is considering changes to its scale and form. It is also considering incentives or priority for the finalists that join the frontier AI program, alongside an AI adoption track aimed at expanding the use of sovereign AI models across industries and the public sector.
An industry official said the government’s two-track approach could have both benefits and challenges for participating companies.
“Running the government-led projects (in succession) could create synergies, but it could also be confusing for participating companies as they consider how to take part in each program,” the official said.